Let $X$ be a set. Recall that an atlas of Banach Manifold on $X$ is a set of charts $\mathcal{A} = \{(U_\alpha, \varphi_\alpha, E_\alpha)\}_\alpha$ where $\varphi_\alpha: U_\alpha\to \varphi_\alpha(U_\alpha)\subset E_\alpha$ is a bijection, $E_\alpha$ is a Banach space, and a compatibility condition between the charts is satisified. This standard definition can be found in the Wikipedia article on Banach Manifolds.
One could so define a Banach Manifold to be a couple $(X, \mathcal{A})$. But a larger atlas is needed to do differential geometry. Namely, all charts that are compatible with $\mathcal{A}$ are also charts. Said differently, if $\mathcal{B}$ is an atlas on $X$ such that $\mathcal{A}\bigcup\mathcal{B}$ is also an atlas on $X$, the two atlas are said to be compatible. So charts of $\mathcal{B}$ are also wanted charts, more precisely $(X, \mathcal{A})$ and $(X, \mathcal{B})$ describe the same manifold.
The collection of all atlases on $X$ is not a set because one can choose a Banach space from the class of all Banach spaces... On Wikipedia, I quote:
Compatibility defines an equivalence relation on the class of all possible atlases on $X$.
Is it possible to define an equivalence relation like that on a proper class? If so, it's not with ZFC, so is it with NBG? Could you give me more details?
Then, still on Wikipedia,
A $C^r$-manifold structure on $X$ is then defined to be a choice of equivalence class of atlases on $X$ of class $C^r$
So let's call $\bar{\mathcal{A}}$ this equivalence class. Now could a Banach manifold be the couple $(X, \bar{\mathcal{A}})$?. I doubt that it's logically consistent to talk about the couple $(X, \bar{\mathcal{A}})$ since $\bar{\mathcal{A}}$ is a proper class. Is it?
I should say that it's not really a question about differential geometry but about logic. Many (most?) of the authors I read about this don't talk about this logical issue. For example, S. Lang's Foundation of Differential Geometry or Bourbaki's "Fascicule des résultats". But Bourbaki seem's to use their own logic.
Finally, I should add that in the case you restrict the possibilities to a set of Banach spaces (for instance $\{\mathbb{R}^n,\, n\geq0\}$), you don't have any issues. But it's more satisfactory to allow every Banach space or classes like Hilbert spaces or finite dimensional normed spaces.
Thanks, Paul.